Explore · 每周精选
发现最受关注的研究论文,追踪研究趋势,订阅感兴趣的期刊与关键词。
Anton Razzhigaev, Matvey Mikhalchuk, Temurbek Rahmatullaev +4 authors
Analysis of Large Language Models reveals that even minor tokens are crucial for context, and introduces LLM-Microscope for assessing token-level nonlinearity and contextual memory.
50 篇论文 · 按点赞排序
Xun Liang, Jiawei Yang, Yezhaohui Wang +9 authors
SurveyX enhances automated survey generation through innovative preprocessing and post-processing methods, improving content and citation quality beyond existing systems.
Wei Xiong, Hanning Zhang, Chenlu Ye +3 authors
Self-rewarding reasoning large language models independently generate and correct their outputs during inference using a two-stage algorithmic framework, enhancing performance without external feedback.
Xiangpeng Yang, Linchao Zhu, Hehe Fan +1 authors
VideoGrain uses space-time attention mechanisms to enhance fine-grained video editing by improving text-to-region control and feature separation.
Yuxiang Wei, Olivier Duchenne, Jade Copet +6 authors
SWE-RL leverages reinforcement learning to enhance large language models' software engineering reasoning, achieving state-of-the-art performance and generalized reasoning skills by learning from extensive open-source software evolution data.
Xiangyu Zhao, Shengyuan Ding, Zicheng Zhang +10 authors
OmniAlign-V improves human preference alignment in MLLMs through supervised fine-tuning and preference optimization without compromising performance on VQA benchmarks.
Xiaoran Liu, Ruixiao Li, Mianqiu Huang +10 authors
The survey examines the advancements and challenges in long-context Large Language Models (LLMs), exploring the architecture, infrastructure, training, and evaluation technologies needed to extend their context length and address the inherent trade-offs.
Gallil Maimon, Avishai Elmakies, Yossi Adi
A method for training high-quality Speech Language Models on a single GPU within 24 hours, demonstrating cost-effectiveness and outperforming predicted computational scaling laws.
Alexander Groshev, Anastasiia Iashchenko, Pavel Paramonov +2 authors
GHOST 2.0, consisting of an Aligner and Blender module, achieves state-of-the-art results in head swapping by preserving identity information, handling extreme poses, and seamlessly integrating the reenacted head into the target background.
Kanana LLM Team, Yunju Bak, Hojin Lee +26 authors
Kanana, a series of bilingual language models, achieves superior performance in Korean and competitive performance in English with lower computational costs through efficient pre-training and post-training techniques.
Jiazhen Pan, Che Liu, Junde Wu +6 authors
MedVLM-R1 enhances medical image analysis through explicit natural language reasoning generation, achieving superior accuracy with limited data and parameters via reinforcement learning rather than supervised fine-tuning.
Jintao Zhang, Chendong Xiang, Haofeng Huang +4 authors
SpargeAttn is a universal sparse and quantized attention method that accelerates diverse models by predicting the attention map and using an online softmax-aware filter to skip unnecessary matrix multiplications.
Juraj Gottweis, Wei-Hung Weng, Alexander Daryin +31 authors
A multi-agent AI system named AI co-scientist aids in scientific discovery by generating and validating novel hypotheses across biomedical areas, demonstrating potential improvements in drug repurposing, target discovery, and bacterial evolution understanding.
Canyu Zhao, Mingyu Liu, Huanyi Zheng +5 authors
DICEPTION, a text-to-image diffusion model, achieves state-of-the-art performance in multiple perception tasks using minimal data and computational resources by leveraging color encoding and conditional image generation.
Max Ku, Thomas Chong, Jonathan Leung +3 authors
TheoremExplainAgent uses agentic planning and Manim animations to generate long-form visual explanations for theorems, revealing deeper reasoning flaws not apparent in text-based explanations.
Zhongyang Li, Ziyue Li, Tianyi Zhou
A test-time re-routing method improves the performance of multimodal models by optimizing routing weights for non-language modalities, enhancing accuracy in challenging tasks without training.
Dongki Kim, Wonbin Lee, Sung Ju Hwang
Mol-LLaMA, a multi-modal instruction-tuned large molecular language model, integrates various molecular encoders to understand general features and provide detailed responses for molecular analysis tasks.
Shijie Huang, Yiren Song, Yuxuan Zhang +4 authors
PhotoDoodle is an image editing framework that uses a two-stage训练 strategy and EditLoRA for realistic blending and perspective alignment, enabling the overlay of decorative elements on photographs with preserved background and artist-specific styles.
Lola Le Breton, Quentin Fournier, Mariam El Mezouar +1 authors
NeoBERT is a next-generation encoder that integrates modern advancements to surpass performance of BERT and RoBERTa on MTEB with a compact parameter footprint.
Tianrui Zhu, Shiyi Zhang, Jiawei Shao +1 authors
KV-Edit uses the KV cache in DiT-based models to preserve background tokens, enabling high-quality image editing with seamless background consistency and low memory usage.
Jingcheng Ni, Yuxin Guo, Yichen Liu +3 authors
A generalized driving world model improves long-horizon and multi-view predictions by integrating masks with diffusion-based feature learning.
Ning Shang, Li Lyna Zhang, Siyuan Wang +5 authors
LongRoPE2 enhances the context window of pre-trained LLMs using rescaled RoPE and mixed context training, maintaining short-context performance with fewer tokens.
Liumeng Xue, Ziya Zhou, Jiahao Pan +19 authors
A new large-scale dataset, Audio-FLAN, supports unified audio-language models by covering diverse tasks across speech, music, and sound for both understanding and generation.
Yifan Pu, Yiming Zhao, Zhicong Tang +14 authors
The Anonymous Region Transformer (ART) enables efficient generation of multi-layer transparent images using text prompts and anonymous region layouts, outperforming full attention methods in speed and layer conflict reduction.
Jianshu Zhang, Dongyu Yao, Renjie Pi +3 authors
VLM$^2$-Bench evaluates vision-language models' ability to link visual cues across images, highlighting significant performance gaps and suggesting improvements to enhance visual capabilities and integration of language-based reasoning.
Chenghao Fan, Zhenyi Lu, Sichen Liu +4 authors
GOAT, a framework that integrates relevant priors using an SVD-structured Mixture-of-Experts and derives a theoretical scaling factor, enhances LoRA MoE's efficiency and performance, approaching Full Fine-Tuning results.
Tushar Aggarwal, Kumar Tanmay, Ayush Agrawal +3 authors
Multilingual language models show inconsistent factual knowledge recall and transferability across languages, highlighting the need for improved cross-lingual generalization.
Xueqing Peng, Triantafillos Papadopoulos, Efstathia Soufleri +7 authors
The first Greek Financial Evaluation Benchmark (Plutus-ben) and Greek Financial LLM (Plutus-8B) address the challenges of financial NLP in Greek due to linguistic complexity and lack of domain-specific data.
Zihao Zeng, Xuyao Huang, Boxiu Li +1 authors
A novel post-training approach, Stick to the Facts (SIFT), improves the reasoning accuracy of large language models by generating and refining stickers, leading to enhanced performance on diverse benchmarks.
Chen-Wei Chang, Cheng-De Fan, Chia-Che Chang +4 authors
GCC utilizes diffusion models to robustly estimate illumination by inpainting color checkers, achieving state-of-the-art results in cross-camera evaluations without sensor-specific training.
北京市昌平区探索星信息技术及软件开发工作室
京ICP备2026059466号